Performance in R using microbenchmark

Basic R

Cast dates

start_date <- "2017-01-01"
end_date <- "2019-12-31"
f1<-function(d2, d1){
  n_weeks <-  floor(as.numeric(difftime(d2, d1, units="weeks")))
}
f2<-function(d2, d1){
  n_weeks <- floor(as.numeric(difftime(as.Date(d2)
    , as.Date(d1), units = "weeks")))
}
m1<-microbenchmark(
  Nocast = f1(end_date, start_date),
  Cast = f2(end_date, start_date),
  times = 1000
)
print(m1)

## Unit: microseconds
##    expr     min       lq     mean   median      uq      max neval
##  Nocast 324.846 334.0790 353.2882 337.0495 348.706 3455.632  1000
##    Cast 116.097 119.7185 129.7460 121.1770 123.936 3095.600  1000

fbox_plot(m1, "microseconds")

Explicit vector length vector(“type”, length) is faster than an empty vector c()

no_size <- function (n){
  x <- c()
  for (i in seq(n)) {
    x <- c(x, i)
  }
}
explicit_size <- function (n){
  x <- vector("integer", n)
  for (i in seq(n)) {
    x[i] <- i
  }
}
m3 <- microbenchmark(
  no_size = no_size(1e4),
  explicit_size = explicit_size(1e4),
  times = 10
)
print(m3)

## Unit: microseconds
##           expr      min        lq       mean     median        uq       max
##        no_size 68590.98 71457.882 72691.4349 72164.6660 73376.905 81942.750
##  explicit_size   353.86   356.136   672.4449   357.6775   370.362  3465.972
##  neval
##     10
##     10

fbox_plot(m3, "microseconds")

which function is slow for some simple situations

vector <- runif(1e8)
w1 <- function(x){
  d <- length(which(x > .5))
}
w2 <- function(x){
  d <- sum(x > .5)
}

m4 <- microbenchmark(
  which = w1(vector),
  nowhich = w2(vector),
  times = 10
)
print(m4)

## Unit: milliseconds
##     expr      min       lq     mean   median       uq      max neval
##    which 635.0059 638.6475 670.2781 644.1300 653.8114 830.0850    10
##  nowhich 223.3879 225.5231 240.4974 231.7384 239.6573 326.1463    10

fbox_plot(m4, "miliseconds")

Column operation is faster than row operation

n <- 1e4
dt <- data.table(
  a = seq(n), b = runif(n)
)
v1 <- function(dt){
  d <- mean(dt[dt$b > .5, ]$a)
}
v2 <- function(dt){
  d <- mean(dt$a[dt$b > .5])
}
m5 <- microbenchmark(
  row_operation = v1(dt),
  column_operation = v2(dt),
  times = 10
)
print(m5)

## Unit: microseconds
##              expr     min      lq      mean   median      uq      max neval
##     row_operation 190.566 199.413 1136.0205 214.1250 321.651 5874.159    10
##  column_operation  74.338  77.264  362.1413  89.3025 126.296 2687.027    10

fbox_plot(m5, "microseconds")

Sequences function safer than 1:n

The function seq prevents when the second part of the 1:x is zero

num <- 1e7
s1 <- function(num){
  d <- mean(1:num)
}
s2 <- function(num){
  d <- mean(seq(num))
}
m6<-microbenchmark(
  noseq = s1(num),
  seq = s2(num),
  times = 30
)
print(m6)

## Unit: milliseconds
##   expr      min       lq     mean   median       uq      max neval
##  noseq 69.82843 69.92892 70.83228 69.98857 70.51530 74.16076    30
##    seq 69.87508 69.97086 71.01114 70.04180 72.05789 74.46660    30

fbox_plot(m6, "miliseconds")

paste0 is faster than glue

large_dataset <- data.table(
  id = 1:1000000,
  value = sample(letters, 1000000, replace = TRUE)
)
a1 <- function(x){
  d <- x |> mutate(code = paste0(id, "_", value))
}
a2 <- function(x){
  d <- x |> mutate(code = glue("{id}_{value}"))
}
m7 <- microbenchmark(
  with_paste = a1(large_dataset),
  with_glue = a2(large_dataset),
  times = 20
)
print(m7)

## Unit: milliseconds
##        expr      min       lq     mean   median       uq       max neval
##  with_paste 586.4241 598.9025 634.3884 607.8835 615.2178 1158.0133    20
##   with_glue 614.6962 620.3687 627.2505 623.5681 631.6961  653.8361    20

fbox_plot(m7, "miliseconds")

for loop vs lapply

# Create a large list
big_list <- replicate(1e5, rnorm(10), simplify = FALSE)

m8 <- microbenchmark(
  lapply = lapply(big_list, mean),
  for_loop = {
    result <- list()
    for (i in seq_along(big_list)) {
      result[[i]] <- mean(big_list[[i]])
    }
  },
  times = 10
)

print(m8)

## Unit: milliseconds
##      expr      min       lq     mean   median       uq      max neval
##    lapply 308.5040 326.8918 356.7680 333.3494 379.6477 489.7449    10
##  for_loop 334.2977 344.8711 379.9272 362.1968 383.8401 485.6251    10

fbox_plot(m8, "miliseconds")

data.table package functions

Date vs IDate

dt <- data.table(
  Date = as.Date('2023-01-01') + 0:99999,
  iDate = as.IDate('2023-01-01') + 0:99999,
  Value = rnorm(100000)
)

nd1 <- as.Date('2023-01-01')
nd2 <- as.Date('2023-01-10')
id1 <- as.IDate('2023-01-01')
id2 <- as.IDate('2023-01-10')

date_between_operation <- function(nd1, nd2) {
  dt |> filter(Date >= nd1 & Date <= nd2)
}
idate_between_operation <- function(id1, id2) {
  dt |> _[data.table::between(iDate, id1, id2)]
}

m9 <- microbenchmark(
  Date = date_between_operation(nd1, nd2),
  iDate = idate_between_operation(id1, id2),
  times = 200L
)
print(m9)

## Unit: microseconds
##   expr      min        lq      mean   median        uq      max neval
##   Date 1464.836 1736.2070 1972.7392 1791.870 2157.2935 3758.328   200
##  iDate  568.812  596.6795  686.3415  615.369  725.0735 2158.130   200

fbox_plot(m9, "miliseconds")

Base R switch vs Dplyr case_when (for simple tasks)

switch_function <- function(x) {
  switch(x,
         "a" = "apple",
         "b" = "banana",
         "c" = "cherry",
         "default")
}
case_when_function <- function(x) {
  case_when(
    x == "a" ~ "apple",
    x == "b" ~ "banana",
    x == "c" ~ "cherry",
    TRUE ~ "default"
  )
}
# Create a vector of test values
test_values <- sample(c("a", "b", "c", "d"), 1000, replace = TRUE)
m10 <- microbenchmark(
  switch = sapply(test_values, switch_function),
  case_when = sapply(test_values, case_when_function),
  times = 200L
)
print(m10)

## Unit: microseconds
##       expr        min         lq       mean      median         uq        max
##     switch    636.518    645.806    668.048    651.7675    660.253   2045.089
##  case_when 215659.545 227684.179 231731.279 231078.8330 234194.967 371896.744
##  neval
##    200
##    200

fbox_plot(m10, "microseconds")

data.table fcase vs Dplyr case_when

set.seed(123)
n <- 1e6
data <- data.table(
  id = seq(n),
  value = sample(seq(100), n, replace = TRUE)
)

casewhenf <- function(data){
  df <- data |> 
    mutate(category = case_when(
      value <= 20 ~ "Low",
      value <= 70 ~ "Medium",
      value > 70 ~ "High"))
}
fcasef <- function(data){
  df <- data |> 
    mutate(category = fcase(
      value <= 20, "Low",
      value <= 70, "Medium",
      value > 70, "High"))
}
m11 <- microbenchmark(
  case_when = casewhenf(data),
  fcase = fcasef(data),
  times = 20
)
print(m11)

## Unit: milliseconds
##       expr      min       lq     mean   median       uq      max neval
##  case_when 56.63917 57.74803 62.79273 60.70106 63.26526 85.27853    20
##      fcase 20.56803 20.78007 22.06972 20.86613 22.44801 27.40792    20

fbox_plot(m11, "miliseconds")

data.table fcoalesce vs tidyr replace_na

set.seed(123)
DT <- data.table(
  ID = 1:1e6,
  Value1 = sample(c(NA, 1:100), 1e6, replace = TRUE),
  Value2 = sample(c(NA, 101:200), 1e6, replace = TRUE)
)

# Define the functions
replace_na_f <- function(data){
  DF <- data |> 
    mutate(Value1 = replace_na(Value1, 0),
           Value2 = replace_na(Value2, 0)) |> 
    as.data.table()
}
fcoalesce_f <- function(data){
  DF <- data |> 
    mutate(Value1 = fcoalesce(Value1, 0L),
           Value2 = fcoalesce(Value2, 0L))
}
m12 <- microbenchmark(
  treplace_na = replace_na_f(DT),
  tfcoalesce = fcoalesce_f(DT),
  times = 20
)
print(m12)

## Unit: milliseconds
##         expr      min       lq     mean   median       uq       max neval
##  treplace_na 7.696541 7.917915 8.392480 8.240813 8.383318 10.932434    20
##   tfcoalesce 1.587544 1.801179 2.517975 2.116061 3.430525  4.121957    20

fbox_plot(m12, "miliseconds")

data.table notation vs dplyr notation

dt <- data.table(field_name = c("argentina.blue.man.watch", 
                                "brazil.red.woman.shoes", 
                                "canada.green.kid.hat", 
                                "denmark.red.man.shirt"))

# Filter rows where 'field_name' does not contain 'red'
dtnot <- function(data){
  filtered_dt <- data |> _[!grepl("red", field_name)]
}
dplyrnot <- function(data){
  filtered_dt <- data |> filter(!grepl("red", field_name))
}

m13 <- microbenchmark(
  data_table_not = dtnot(dt),
  dplyrnot = dplyrnot(dt),
  times = 100
)
print(m13)

## Unit: microseconds
##            expr     min       lq     mean   median      uq      max neval
##  data_table_not  98.965 111.6635 159.9027 134.4665 160.565 2078.652   100
##        dplyrnot 649.944 670.0660 778.6597 688.2800 737.452 2724.448   100

fbox_plot(m13, "microseconds")

data.table melt vs tidyr pivot_longer

large_data <- data.table(
  id = 1:100000,
  var1 = rnorm(100000),
  var2 = rnorm(100000),
  var3 = rnorm(100000),
  var4 = rnorm(100000)
)
# Benchmarking
m14 <- microbenchmark(
  tidyr_pivot_longer = {
    long_data_tidyr <- pivot_longer(large_data, cols = starts_with("var"), 
                                    names_to = "variable", values_to = "value")
  },
  data_table_melt = {
    long_data_dt <- melt(large_data, id.vars = "id", variable.name = "variable", 
                         value.name = "value")
  },
  times = 10
)

print(m14)

## Unit: microseconds
##                expr      min       lq      mean   median       uq       max
##  tidyr_pivot_longer 6311.746 6373.201 8997.8637 6467.141 6840.143 30905.990
##     data_table_melt  444.791  484.875  581.3066  499.859  703.734   855.367
##  neval
##     10
##     10

fbox_plot(m14, "microseconds")

data.table CJ vs tidyr expand_grid

vec1 <- seq(1000)
vec2 <- seq(1000)

# Define functions to be benchmarked
expand_grid_func <- function() {
  return(expand_grid(vec1, vec2))
}

CJ_func <- function() {
  return(CJ(vec1, vec2))
}

# Perform benchmarking
m15 <- microbenchmark(
  expand_grid = expand_grid_func(),
  CJ = CJ_func(),
  times = 10
)

print(m15)

## Unit: microseconds
##         expr      min       lq     mean   median       uq      max neval
##  expand_grid 2271.882 2356.501 2764.077 2474.796 2992.037 4132.696    10
##           CJ  458.717  482.571  776.949  598.964  919.878 1913.733    10

fbox_plot(m15, "microseconds")

data.table rbindlist vs R rbind

# Sample data
size = 1e4
set.seed(44)
df_list <- replicate(50, data.table(id = sample(seq(size), size, replace = T),
                                    value = rnorm(size)), simplify = F)

simple_bind <- function(list_of_dfs){
  do.call(rbind, list_of_dfs)
}

dplyr_bind <- function(list_of_dfs){
  bind_rows(list_of_dfs)
}

dt_bind <- function(list_of_dfs){
  rbindlist(list_of_dfs, fill = F)
}

# Benchmark both methods
m16 <- microbenchmark(
  dt_ver = dt_bind(df_list),
  simple = simple_bind(df_list),
  dplyr_ver = dplyr_bind(df_list),
  times = 30
)

print(m16)

## Unit: microseconds
##       expr       min        lq       mean     median        uq       max neval
##     dt_ver   418.913   510.543   666.5973   554.6205   651.026  2054.636    30
##     simple   485.848   521.012   676.3720   577.7635   628.263  2173.098    30
##  dplyr_ver 10111.641 10247.625 10708.0204 10358.0360 11069.479 12881.703    30

fbox_plot(m16, "microseconds")

stringr word vs tidyr separate vs data.table tstrsplit

set.seed(123)
n <- 1e4
df <- data.table(text = paste("word1", "word2", "word3", "word4", "word5", sep = "."), stringsAsFactors = F)
df <- df[rep(1, n), , drop = F]

# Using tidyr::separate
separate_words <- function() {
  df |> 
    separate(text, into = c("w1", "w2", "w3", "w4", "w5"), sep = "\\.", remove = F) |> 
    select(-c(w1, w2, w4))
}

# Using stringr::word
stringr_words <- function() {
  df |> 
    mutate(
      w3 = word(text, 3, sep = fixed(".")),
      w5 = word(text, 5, sep = fixed("."))
    )
}

datatable_words <- function() {
  df |> _[, c("w3", "w5") := tstrsplit(text, "\\.")[c(3, 5)]]
}

m17 <- microbenchmark(
  separate = separate_words(),
  stringr = stringr_words(),
  dt = datatable_words(),
  times = 10
)

print(m17)

## Unit: milliseconds
##      expr       min        lq      mean    median        uq       max neval
##  separate  77.47755  80.21303  95.79084  90.54606  93.56122 179.10341    10
##   stringr 184.64272 186.57560 192.32802 188.28115 201.82303 205.98428    10
##        dt  12.51353  12.56643  13.04390  12.63552  13.03781  15.25742    10

fbox_plot(m17, "miliseconds")

data.table na_omit vs dplyr drop_na

# Sample data
set.seed(123)
n <- 1e6
df <- data.table(
  x = rnorm(n),
  y = sample(c(NA, 1:100), n, replace = TRUE),
  z = sample(c(NA, letters), n, replace = TRUE),
  stringsAsFactors = F
)

# Benchmark both methods
m18 <- microbenchmark(
  dplyr_drop_na = {
    df |> drop_na()
  },
  data_table_na_omit = {
    dt |> na.omit()
  },
  times = 10
)

print(m18)

## Unit: microseconds
##                expr      min       lq      mean    median       uq      max
##       dplyr_drop_na 9368.785 9443.985 9511.2784 9455.4560 9506.521 9960.990
##  data_table_na_omit    8.576    9.478   51.2127   55.0025   62.867  177.271
##  neval
##     10
##     10

fbox_plot(m18, "microseconds")

Parallel processing

lapply vs parallel mclapply

# Sample data
set.seed(123)

size = 1e4
n_cores = parallelly::availableCores()

df_list <- replicate(100, data.table(id = sample(seq(size), size, replace = T),
                                    value = rnorm(size)), simplify = F)
extra_df <- data.table(id = sample(seq(size), size, replace = T), 
                       extra_value = runif(size))

# Sequential join
sequential_join <- function() {
  lapply(df_list, function(df) {
    merge(df, extra_df, by = "id", allow.cartesian = T)
  })
}

# Parallel join using mclapply
parallel_join <- function() {
  mclapply(df_list, function(df) {
    merge(df, extra_df, by = "id", allow.cartesian = T)
  }, mc.cores = n_cores, mc.silent = T, mc.cleanup = T)
}

# Benchmark both methods
m19 <- microbenchmark(
  sequential = sequential_join(),
  parallel = parallel_join(),
  times = 10
)

print(m19)

## Unit: milliseconds
##        expr      min       lq     mean   median       uq      max neval
##  sequential 299.3829 309.7949 329.8663 317.0577 348.9830 386.8511    10
##    parallel 151.8398 158.9782 169.3643 164.5133 175.4441 193.8522    10

fbox_plot(m19, "miliseconds")

Duckdb

Dockdb files vs parquet

with_parquet <- function(){
  fp_data <- "/conf/posit_azure_logs/data"
  
  data_1 <- open_dataset(file.path(glue::glue("{fp_data}/golden_data_in_progress"))) |>
    select(
      date, hours, time,
      ALL_WIP_CP_day_session, ALL_WIP_CP_night_session,
      ALL_WIP_BP_day_session, ALL_WIP_BP_night_session,
      ALL_WIP_CP_DS_mem_limit, ALL_WIP_CP_NS_mem_limit,
      ALL_WIP_BP_DS_mem_limit, ALL_WIP_BP_NS_mem_limit,
      ALL_WIP_CP_DS_mem_request, ALL_WIP_CP_NS_mem_request,
      ALL_WIP_BP_DS_mem_request, ALL_WIP_BP_NS_mem_request,
      ALL_WIP_CP_DS_mem_max, ALL_WIP_CP_NS_mem_max,
      ALL_WIP_BP_DS_mem_max, ALL_WIP_BP_NS_mem_max,
      ALL_WIP_CP_node_total, ALL_WIP_BP_node_total
    ) |>
    mutate(
      computepool_node_mem = ALL_WIP_CP_node_total * (160 * 1024),
      bigpool_node_mem = ALL_WIP_BP_node_total * (256 * 1024),
      ALL_WIP_day_session = ALL_WIP_CP_day_session + ALL_WIP_BP_day_session,
      ALL_WIP_night_session = ALL_WIP_CP_night_session + ALL_WIP_BP_night_session,
      ALL_WIP_node_total = ALL_WIP_CP_node_total + ALL_WIP_BP_node_total,
      total_mem_limit = ALL_WIP_CP_DS_mem_limit + ALL_WIP_CP_NS_mem_limit + ALL_WIP_BP_DS_mem_limit + ALL_WIP_BP_NS_mem_limit,
      total_mem_request = ALL_WIP_CP_DS_mem_request + ALL_WIP_CP_NS_mem_request + ALL_WIP_BP_DS_mem_request + ALL_WIP_BP_NS_mem_request,
      total_mem_max = ALL_WIP_CP_DS_mem_max + ALL_WIP_CP_NS_mem_max + ALL_WIP_BP_DS_mem_max + ALL_WIP_BP_NS_mem_max,
      total_node_mem = computepool_node_mem + bigpool_node_mem,
      average_session_per_node = ifelse(ALL_WIP_node_total != 0,
                                         (ALL_WIP_day_session + ALL_WIP_night_session) / ALL_WIP_node_total, 0)
    ) |>
    collect() |>
    as.data.table()
}

with_duckfile <- function(){
  file.copy("/conf/posit_azure_logs/gatzos01/gd_inprogress.duckdb", "gd_inprogress.duckdb")
  
  data_2 <- res_duckdb_sql <- dbGetQuery(
    conn = dbConnect(duckdb::duckdb(), dbdir = "./gd_inprogress.duckdb"),
    statement = glue("select date, hours, time,
        ALL_WIP_CP_day_session, ALL_WIP_CP_night_session,
        ALL_WIP_BP_day_session, ALL_WIP_BP_night_session,
        ALL_WIP_CP_DS_mem_limit, ALL_WIP_CP_NS_mem_limit,
        ALL_WIP_BP_DS_mem_limit, ALL_WIP_BP_NS_mem_limit,
        ALL_WIP_CP_DS_mem_request, ALL_WIP_CP_NS_mem_request,
        ALL_WIP_BP_DS_mem_request, ALL_WIP_BP_NS_mem_request,
        ALL_WIP_CP_DS_mem_max, ALL_WIP_CP_NS_mem_max,
        ALL_WIP_BP_DS_mem_max, ALL_WIP_BP_NS_mem_max,
        ALL_WIP_CP_node_total, ALL_WIP_BP_node_total,
        ALL_WIP_CP_node_total * 160 * 1024 as computepool_node_mem,
        ALL_WIP_BP_node_total * 256 * 1024 as bigpool_node_mem,
        ALL_WIP_CP_day_session + ALL_WIP_BP_day_session as ALL_WIP_day_session,
        ALL_WIP_CP_night_session + ALL_WIP_BP_night_session as ALL_WIP_night_session,
        ALL_WIP_CP_node_total + ALL_WIP_BP_node_total as ALL_WIP_node_total,
        ALL_WIP_CP_DS_mem_limit + ALL_WIP_CP_NS_mem_limit + ALL_WIP_BP_DS_mem_limit + ALL_WIP_BP_NS_mem_limit as total_mem_limit,
        ALL_WIP_CP_DS_mem_request + ALL_WIP_CP_NS_mem_request + ALL_WIP_BP_DS_mem_request + ALL_WIP_BP_NS_mem_request as total_mem_request,
        ALL_WIP_CP_DS_mem_max + ALL_WIP_CP_NS_mem_max + ALL_WIP_BP_DS_mem_max + ALL_WIP_BP_NS_mem_max as total_mem_max,
        computepool_node_mem + bigpool_node_mem as total_node_mem,
        CASE 
          WHEN ALL_WIP_node_total != 0 THEN (ALL_WIP_day_session + ALL_WIP_night_session) / ALL_WIP_node_total
          ELSE 0
        END AS average_session_per_node
      from gdinprog"),
    immediate = TRUE) |> 
    as.data.table()
  
  file.remove("./gd_inprogress.duckdb")
}

m21 <- microbenchmark(
  with_parquet = with_parquet(),
  with_duckfile = with_duckfile(),
  times = 3
)

print(m21)

## Unit: milliseconds
##           expr        min         lq       mean    median         uq        max
##   with_parquet 12636.6832 12857.6282 16831.3121 13078.573 18928.6265 24778.6797
##  with_duckfile   464.8961   541.4405   607.0797   617.985   678.1714   738.3579
##  neval
##      3
##      3

fbox_plot(m21, "miliseconds")